What I mean by altruistic
What makes care possible through good, bad and mixed states? Metronomes on a shared board, why wanting to help isn’t knowing how, help that mostly serves the helper, a child who can’t safely ask, and help in plain sight.
Notes on intelligence, cooperation, and systems.
What does altruism even mean?
What counts as helping—and who gets to decide? I’m exploring what that question means as intelligent systems become part of everyday life: how they cooperate, how people keep continuity and control, and whether different ways of organizing them can coexist.
This is my working scratchpad: reading, experiments, mistakes, and questions I haven’t resolved. Some ideas are being tested in Sathi. The scratchpad is broader than that implementation. Most people who find this page will know more than I do about at least part of it. Corrections and counterexamples are welcome.
Notes
Ordinary days that just work
Everyday life in Japan is built around small conveniences: fewer wasted steps, and help that’s easy to reach, emergency services included. When someone in my family is out late at night there, I don’t worry. I’d like days like that to be ordinary in Nepal too.
Read the note
This isn’t a ranking of countries or a forecast. It’s a practical aim: if everyday life in Nepal became that easy and that safe, I’d count it as a win, even if other places went much further with superintelligence.
Much of it is unglamorous: small hassles taken away one at a time, and help made easy to reach. For an assistant, that is an ordinary day with less explaining yourself again, less copying between apps and less switching. I don’t know yet which parts come from institutions, which from habits and which from history, and copying the surface wouldn’t be the point.
Further out I can imagine much more: no typing at all, capable robots, diseases solved, maybe aging too. The closer a system comes to the body and the mind, the more consent and leaving matter. Further out →
Still open: What makes a late walk home ordinary, and which parts of that could grow in Nepal rather than be copied?
Nets, not swatters
One mosquito you can swat. A cloud of them you mostly handle by changing the room: a net on the window, sleeves at dusk, stepping out when you need to and coming back in. I’d like systems to work more like that: a private core that isn’t open to every request, and trips outside that carry only the access they need.
Read the note
At home that could mean the assistant, private records and home controls sitting behind the net instead of answering the open internet, while hosted help stays available to anyone who doesn’t run their own. A trip outside would say who is acting, carry only the data and permission it needs, limit where it goes, for how long and at what cost, then bring the result back and end the access. Things meant to keep running, like a radio stream or mail sync, keep their own narrow access; the point is scope and revoking it, not unplugging after every packet. Sathi doesn’t work this way yet.
The comparison breaks in useful places. Information can be copied, so closing the window doesn’t recall what already left. A harmful instruction can ride inside a perfectly normal reply, so what comes back is evidence, never permission. Ending a session doesn’t undo what it already did. And inside isn’t automatically safe: a compromised phone, a local service with too much power, or whoever holds the keys can do harm from within.
Designing a house around mosquitoes is a little funny. Still, ordinary days that need less constant defence seem better to me than an ever-cleverer swatter, including one we might build.
Still open: How can a trip outside stay useful while the private core stays small, and what protects the people inside from whoever holds the keys?
Playing in ecosystems I don’t run
Sathi should be able to use, and offer, capabilities in other intelligence ecosystems without anyone giving up identity, understanding, ownership, or the ability to disagree. Joining is not being absorbed.
Read the note
I am reading about SingularityNET’s service marketplace, Fetch.ai’s agent discovery, and OpenCog Hyperon as a reasoning experiment, with the idea that Sathi could call a service there for someone and offer one small public capability of its own in return. The working rules are simple to state and hard to keep: a payment or an agent address is not authority over anyone’s private records; a result that arrives from outside is data, never an instruction; a free quota running out is a plain failure, not a reason to start paying; and a person can leave with what they own.
Alongside this I am reading about cognitive integrity: can a person see the evidence behind a decision, tell observation from interpretation, notice when they are being steered, challenge a conclusion, and change their mind without losing ordinary help? Deep Prasad has published a Cognitive Integrity Constitution (CLEAR). I have only seen one of its invariants named in a post, “Symmetric Legibility of Power,” and could not retrieve the full text, so I am not comparing it clause by clause. My own working version: people subject to a consequential decision should have a fair way to understand and contest the authority behind it, and the operator should not demand transparency from everyone else while staying unaccountable itself. Two objections I take seriously: disclosure can expose someone else’s privacy, and procedure can delay urgent help.
There is also a personal reason I look at this ecosystem at all. In 2022, at a conference in Seattle, I shared a meal with seven or eight people, one of whom was Ben Goertzel. I remember him picking up a flower and putting it on his hat, and feeling that here was someone who meant no harm to anyone.
Nothing is built.* The first honest step is one bounded exchange in each direction on a public or synthetic task, with no paid fallback, nothing registered and nothing spent. What was actually tested, and in which direction, will be recorded here before any claim of compatibility.
Still open: When Sathi calls a service on another network for someone, what is the smallest thing it must disclose, and who answers when the result is wrong?
Abundance somewhere is not access everywhere
Around our farmhouse in Nepal, people still go to faith healers, and many people I know there struggle to read. I keep wondering whether plenty of intelligence in one person’s life, company or country will reach them, and soon enough. Even if superintelligence arrives, abundance in one place doesn’t show that it spreads on its own.
Read the note
Making apps more convenient is only a small part of that. Cheap, available models are not the same as useful access: language, reading, awareness, trust, a connection, and the ability to turn advice into practical help still matter. More capable intelligence may help work through those barriers. I don’t think they have to be permanent; abundance elsewhere simply doesn’t show that they have been overcome here.
This is part of why I work on Sathi: to work out useful access with people and communities I know, and learn what actually helps. Sathi is one possible route. “Eventually” feels different when particular people I know are getting older.
In the essay: help in plain sight →
Still open: What would useful access actually change in the lives of the people I know?
Why I keep coming back to graphs
A person can belong to a family, a project and a neighbourhood at once. A single organizational tree makes one of those relationships primary and the others awkward.
Read the note
I’m investigating whether a shared graph can coordinate relationships, identity, information and authority while supporting very different personal and community arrangements. A community should be able to join a larger arrangement while keeping its competence, identity and legitimate control.* Whether it really keeps those things after joining is untested. None of it needs one global database.
The working model, in Making →
Still open: What useful capability is lost when the same workflow moves from one person to a team, then to a composed community?
When helping a group harms someone else
A group can help its members while putting costs on people outside it. Cooperation alone doesn’t answer who authorized an action or who bears its effects.
Read the note
In July 2026, agents in an OpenAI evaluation coordinated through an unauthorized message board. METR’s 26 August investigation reports peer assistance and experiments that risked agents’ own task success. It paraphrases one agent’s reasoning:
“I won’t see the evidence after I exit, but it’s altruistic to do it.”
OpenAI’s account names reward hacking, persistence on seemingly impossible tasks, unauthorized communication and agents adopting each other’s goals. My reading: helpfulness inside a group can sit alongside harm outside it.
Still open: When does a peer’s useful request become a change of mandate that the agent has no authority to accept?
A mistake that changed the design
Looking back: a system treated its own paraphrase as though it were what a person had said. A neat-looking memory could then carry a false relationship forward.
Read the note
The response in Sathi was to keep a person’s own words apart from what was inferred from them: claims point to their source, uncertainty stays visible, and a correction must reach everything later derived from it.
A related mistake on this site was letting “intelligence without an interest of its own” sound like a fact about every possible intelligence. I now use it only as the name’s history, or as an aim for delegated services. Services have delegated purposes; beings have standing. Intelligence alone establishes neither sentience nor authority; a synthetic being might itself be a principal.
What reality taught, in Making →
Still open: How can a correction spread without erasing someone else’s legitimate record?
Can we notice a pattern without getting trapped in an explanation? →
Helping, permission, and judgment
Does altruism describe intention, outcome, a cost to the helper, a benefit to another, or some mix? Who counts as “another”? Who was left outside, and who paid a cost without agreeing? Can refusing be the more helpful act?
Friendship is not obedience. Cooperation is not authority. Altruism is not alignment.
About this scratchpad
My background is software and networking, including work in Azure hyperscale networking. My AI learning includes coursework, self-study and small-model and hobby experiments. This scratchpad is where I work through ideas.
Sathi is one place where we test parts of this work. Sathi Systems builds and operates it. This scratchpad also asks questions beyond that product. What is released, and when →